Machine learning is either the future of medicine or a distraction from clinical medicine – if the lively discussion at ISTH Congress is anything to go by.
The discussion was generated after Dr John Willan, from Oxford University Hospitals, told the Congress that machine learning algorithms might offer a more advanced and nuanced method of DVT risk assessment than current methods.
His team’s artificial neural network was “trained” on data from about 5,000 patients with suspected DVT. The data included all individual components of the Well’s score, D-dimer and ultrasound results.
The subsequent algorithm was then applied to a test set of almost 2,000 patients with suspected DVT.
The algorithm identified a rate of positive DVT in the testing dataset slightly lower than the rate in the training dataset (10.8% v 11.9%) – suggesting a degree of selection bias for lower risk patients.
The proof-of-principle study, previously published in the British Journal of Haematology, also found machine learning could possibly reduce reliance on ultrasound.
It could exclude DVT in 37.5% of patients without the need for ultrasound.
Dr Willan said that the difference with machine learning algorithms was they did not rely on linear relationships between all the variables.
The relative contribution of each variable and the subsequent estimated risk could then be very different in patients who, when scored with traditional risk assessment tools, might appear to have the same level of risk.
Dr Willan, who studied medicine after undergraduate studies in maths and physics, told the limbic machine learning offered a huge opportunity to do good.
“All our historical scoring systems have been designed around ease of use by pen and paper in a clinic. That’s what’s motivated them to work and they’re great.”
“But I just think that now we all have a computer in our pocket and on our desktop, we should be thinking about moving to an algorithm that can potentially incorporate more information and potentially give a more accurate outcome.”
“I think it’s going to be a paradigm shift for clinicians in general but I don’t see why that shift can’t be made. At some point in time it is going to happen.”
Clinicians can experiment with the algorithm at Oxford Neural Network.